TDA-Net:A novel attention-based model with temporal feature fusion for Motor Imagery Classification

Motor imagery (MI) based on electroencephalography (EEG) is a prominent brain-computer interface (BCI) signal that has been widely adopted for applications in medical rehabilitation to help disabled patients. Traditional models based on attention mechanisms have achieved satisfactory results in subject-dependent MI classification. However, most methods solely utilize attention mechanisms to capture long-term dependencies in MI-EEG signals but fail to capture the potential interaction between local and global temporal information, resulting in suboptimal performance. In this paper, we propose a novel attention-based network that integrates a multi-scale pooling layer with the self-attention mechanism to effectively extract local and global temporal features from MI-EEG signals. To further integrate local and global temporal contextual information, we introduce a convolutional fusion module, thereby improving overall decoding performance. The proposed model demonstrates superior performance over the state-of-the-art algorithms in three public datasets (BCIC-IV-2a, BCIC-IV-2b, HGD) with an accuracy of 83.90%, 90.38% and 96.16% for MI subject-dependent tasks, respectively. Ablation study and visualization analysis (t-SNE, Grad-CAM) further validate the effectiveness and interpretability of our model, highlighting its superior decoding performance.